Sequoia Is Right That AI Is Coming For Services, But Three Things Don't Get Talked About Enough
Sequoia recently published a piece called Services: The New Software. The central argument: the next generation of great AI companies won't sell tools to professionals — they'll sell the work those professionals used to do. Don't build the copilot. Build the autopilot.
I think they're right. It's one of the clearest articulations of a real pattern that's already playing out across accounting, legal, insurance, and healthcare billing. The companies winning in those categories aren't selling software seats — they're selling closed books, drafted contracts, filed claims.
I keep coming back to the part the framework doesn't fully address:
Selling outcomes is a fundamentally different business to operate than selling software. And most founders underestimate just how different.
The Liability Gap Nobody Talks About
When you sell software, you sell a tool. The customer uses the tool. The customer owns the outcome.
When you sell a service — a closed set of books, a filed insurance claim, a drafted NDA — you own the outcome. That changes everything about your liability exposure, your error rate tolerance, your customer success model, and how you hire.
Software companies can ship with bugs. The customer finds them, reports them, and you patch. An AI that miscodes a medical bill doesn't get a bug report — it generates a denied claim, a compliance audit, or a lawsuit. An AI that drafts a flawed NDA doesn't get a one-star review — it creates legal exposure for a real business.
The autopilot model pushes accountability downstream from the customer to you. That's the business model. But it also means your error rate has to be dramatically lower than any software product ever needed to be — before you can actually sell the outcome with confidence.
The Distribution Problem Isn't Solved by the Model
Sequoia's opportunity map is organized by labor TAM and intelligence ratio. It's a useful lens for identifying where AI can technically do the work.
But the hardest part of the autopilot model isn't building AI that can do insurance brokerage or medical billing. It's getting a CFO to trust an AI with their compliance obligations, or convincing a business owner to stop calling their broker of 15 years.
Incumbents in services businesses don't just have customers. They have relationships built on years of personal accountability. The broker who handled your claim in 2019. The accountant who called you on a Sunday when something looked wrong. That relationship carries real switching cost — and it's not going to dissolve just because an AI can technically do the same task.
The autopilot companies that win won't just out-AI the incumbents. They'll find the customers who are already dissatisfied with the relationship — the ones whose broker is unresponsive, whose accountant is overloaded, whose outside counsel takes a week to turn a simple NDA. That's where the wedge actually goes in.
The Hybrid Phase Everyone Skips Over
The Sequoia piece draws a clean line between copilots and autopilots. In practice, almost every successful autopilot company starts somewhere in between.
You can't launch as a pure autopilot on day one. Your AI isn't good enough yet. Your error rate isn't low enough. Your liability coverage isn't in place. You haven't built the domain-specific data that makes your model meaningfully better than the base model on the tasks that matter.
So you hire human experts to back-stop the AI. You build a review layer. You staff operations people who handle the edge cases. This is not a failure of the autopilot vision, it is the correct path to get there. But it means your unit economics in the early years look nothing like a software company. Your gross margins are services-business margins, not SaaS margins, until your AI earns the right to operate without a human in the loop.
The real question isn't whether you should build an autopilot. It's whether you can fund the human-backed version long enough for the AI to earn the right to replace it.
What This Means If You're Building in One of These Categories
The Sequoia framework is a useful map. But a map tells you where the territory is valuable — not how to cross it. A few things worth thinking hard about before you go all-in on the autopilot model:
Get specific about your error tolerance before you get customers. What does a mistake actually cost in your category? A wrong insurance recommendation has different consequences than a miscoded medical bill. Know that number before you price your service, before you set your SLA, and before you decide how much human review you need.
Find the customers who are already underserved by the relationship. Don't try to take the happy customer from a great incumbent. Find the mid-market company whose broker is a regional firm with 300 clients, whose accountant has been too busy to return calls since February, whose outside counsel charges $450 an hour for work a good AI can do in seconds. That customer is ready to switch.
Be honest about your gross margin path. If you have humans in the loop today, model explicitly when they come out, what triggers their removal, and what your margins look like at each stage. Investors who understand this model will respect the honesty. The ones who don't are the ones who'll be surprised when your Series B looks like a services business.
The autopilot opportunity is real. The biggest AI companies of this decade probably will look like services firms.
But the path to getting there runs through the unglamorous work of earning customer trust one outcome at a time, building liability infrastructure that most software founders have never thought about, and being honest about the human backstop you need before your AI is ready to fly alone.
Read the full Sequoia piece at sequoiacap.com/article/services-the-new-software. It's worth your time.
